An insight management platform for computer vision model performance. Manot pinpoints where, how, and why computer vision models fail. It accelerates model refinement and redeployment processes by 10x, boosts accuracy by 20%, and reduces costs by 32%.
Manot is a great platform. Product is fulfilling a particular gap in the field with interesting approach. It provides unique solution even for those who st furst glance could see Manot as a competition. Team that us behind it is fantastic.
Manot is capable of shifting the paradigm from post-production surprises to proactive optimization. Already proven to be super helpful. The team behind it deserves a special standing ovation!
It helps me and my data science team immensely since we can monitor and act instead of just testing the model in dev and then getting surprised if it does not perform well enough in prod!
👋 Hi Product Hunt!
I’m Chinar, co-founder and CEO of Manot.
Grab an ice cream, sit back, relax, and let me take you on our adventure! 🍦🚀
It all began back when I was a computer vision (CV) engineer working on cool projects from surveillance to high-tech drones. But here’s the catch - our AI models were awesome in development but not so awesome in the real world 🙈. I saw models with 95% accuracy during testing begin to fail in production, which causes unhappy customers and a lengthy feedback loop between product managers and CV engineers.
So, myself and a few talented friends rolled up our sleeves and got to work. After endless cups of coffee ☕, along with some laughs and cries… Manot was born. We were on a mission - to make CV models smarter before and after being in the real world! 🤓
And guess what? Our little mission got some love! We raised a pre-seed round with the amazing people at Argonautic Ventures, Berkeley SkyDeck, and SmartGateVC 💜.
Being the detectives we are 🕵️🕵️, we talked to over 200 product managers, CV engineers, and data scientists to make sure this problem was felt everywhere. And the response? Mind-blowing 🤯! We’ve got our MVP into the hands of pilot customers, including two Fortune 500 companies!
Here is a quick glimpse into how it works 💻: We developed a scoring algorithm that takes the inference results on a given model’s test dataset, and spits out predictions about the model’s blind spots 🧠. We offer on-premise or cloud solutions along with access to our 5 billion image data lake and generative AI modules. For a deeper dive into our tech, see Erik’s comment below!
Now, here’s the cherry top 🍒: Manot now has a free tier! We’re throwing open the doors so everyone can play with Manot. This isn’t just about growing the platform; it’s about learning from you. So ask questions, reach out, and let’s make something amazing together!
Thank you for being a part of this adventure - your support, feedback, and ideas are what keep us going 🚀💜
Cheers,
Chinar
@edgar_ohanyan2 Thanks! We are use case agnostic! Here are some of the many use cases Manot has addressed: surveillance, drone automation, autonomous vehicles, construction, manufacturing, industrial robots, etc. Happy to walk you through it tailored to your use case!
Report
@chinarmovsisyan big congrats on the launch! Just shot and posted a landing page feedback vid on X 👏
🧑💻Hey Product Hunt,
I’m Erik, the R&D Lead at Manot.
Throughout my career in Computer Vision 💻👁️, I have trained various predictive and generative models, all of which require specific and thorough diagnostics and evaluation before being ready for production. Building these pipelines for every single task and model takes a lot of time and resources ⌛, and even doing all of this does not guarantee the same level of performance on your production data as there can be a significant gap between your test sets and the production data.
Manot is the remedy to these problems. We have created a solution that readily diagnoses classification, segmentation, and detection models without even requiring the model itself, using only a small set of its predictions, ground truths, and raw images. By analyzing the weaknesses of your models we are able to pinpoint scenarios where your model will fail both from your raw data pool, huge data lakes of our partners or even generate such cases by leveraging the state-of-the-art GenAI capabilities 🧠
Leave comments below and we can discuss! 🚀
-
Erik
Report
@ero1311 thank you very much for the product and technical description❤️ love it❤️ wanted to know do you have some lower bound for custom test set or your algo works even with single test set image ?
@tigran_hovhannisyan9 that's a really good question! Actually there’s no lower bound for this, but the more images you provide the better will our solution understand the weaknesses of your model. Considering our numerous experiments across different tasks and datasets we would recommend to start with at least 100 images.
Report
@ero1311 thank you very much! One more question. Does that number changes for different tasks such as classification, detection, etc. or it’s the same for all of them ?
Report
Hello @ero1311,
I'm genuinely impressed by the quality of your work—it's quite remarkable. Reflecting on my past experiences, a particular question arises concerning models designed for predicting highly specific objects that might not be well-represented in public datasets. This situation naturally raises concerns about the scarcity of good data samples. How your algorithms will typically address such circumstances?
Thanks!
@tigran_hovhannisyan9 I would say the number doesn't depend that much on the task at hand, rather on the number of semantic classes you are planning to predict. I mean if you are trying to predict 120 classes starting with 100 images obviously will not be enough for Manot to analyze the weeknesses of your model :)
🤙 Hey there Product Hunters!
I’m Haig, co-founder and CPO of Manot.
Let’s talk about the life of a product manager in AI 🌍💻. For the past 5 years, I’ve had one goal: ensure each AI product (and thus the underlying model) is not just good, but great for our customers. But here’s the thing, time and time again the models prove to be unpredictable.
Every time a model performed poorly or failed, it was back to the whiteboard with the engineering team ✏️. Imagine this process: we detected a problem, reported the problem, wait for a fix, and hold our breath while we hope it works. Then we repeat this process again, and again, and again. It was a never-ending, lengthy feedback loop 👎
That is why Manot is so personal to us. It solves a problem we saw, hated, and did not conquer 🦦. By proactively addressing AI model performance in production, along with a more automated model evaluation and data curation pipeline, we are inherently solving the feedback loop problem.
Excited to hear your thoughts and dive into discussions!
-
Haig
Report
@haig_douzdjian Fellow AI product manager here! - what level of technical knowledge should the average manot PM user have for getting the best out of the product?
Great question @mananahakobyan ! Manot was built with PMs and data teams in mind. The platform requires little technical knowledge. Most of the "work" is done by connecting integrations, which then allows Manot to synergize Product KPIs and Engineering KPIs.
Let me know if you have any more questions!
I think this is a phenomenal platform, I have been in data labeling and AI for more than 3 years now and I think this is one of the tools that will make everyone's time more efficient.
@chinarmovsisyan@ero1311@haig_douzdjian awesomely great job.
P.S. Also Product and ML Engineers finally will be able to speak the same language.
@stefan_radisavljevic Stef jan, thanks for the kind words! We also truly believe our product will save all the time spent on back and forth between Product and ML teams :)
Report
Love the inclusion of product managers in the process. Not equipping them with proper monitoring tools makes the entire product lifecycle rough for everyone involved. Good luck with the launch!
@davit_shadunts great question! Manot is not an Active Learning solution but it covers that:) It provides actionable insights in the form of images pinpointing where, how, and why a model fails … uncovering new categories and samples that contribute to model performance the most.
Report
Congrats on the launch, Manot team. You are tackling a real problem.
I wanted to ask how reliable is your algorithm? How sure are the results for false positives and false negatives?
@garik Great question! We provide you with insights on which the estimated confidence of our scoring function is among the highest. We have evaluated this mechanism indirectly by performing several active learning experiments with both real and publicly available datasets on which our method exhibited significantly better performance than the random baseline and conventional selection mechanisms like entropy-based selections.
Report
@chinarmovsisyan Thank You! Sounds really interesting as state of the art solutions rely on entropy based approach.
Manot